CUDA atomic operations error on compilation

[KlimkinND]

CUDA kernel differentiation with autodiff_deferred fails to compile if the kernel contains a CUDA.@atomic operation. The versions used are: Enzyme v0.13.19, CUDA v5.5.2, Julia v1.11.2. MWE below:

using CUDA
using Enzyme

function test_kern!(y, x)
    i = threadIdx().x

    for j=1:size(y, 1)
        @inbounds CUDA.@atomic y[j] += x[j, i]
    end

    return
end

function grad_test_kern!(y, dy, x, dx)
    autodiff_deferred(Reverse, Const(test_kern!), Const, Duplicated(y, dy), Duplicated(x, dx))

    return
end

d = 10

y = CUDA.zeros(d)
dy = similar(y)

fill!(dy, 1f0)

x = CUDA.randn(d, 128)
dx = zero(x)

@cuda threads=size(x, 2) grad_test_kern!(y, dy, x, dx);

Array(dx)

The resulting error reads:

LLVM error: Cannot select: 0x74f5c70: f32,ch = AtomicLoad<(load acquire (s32) from %ir."'ipc20_unwrap.i", addrspace 1)> 0x67a6a00, 0x74f59d0,